The Questioning Model as an Ethics Pre-Phase

Guiding question: How does structured question expansion change the decision before an AI system is selected, tested, or introduced?

An apparently clear question can already prejudge the decision

A small spare-parts company is considering using a language model to create summaries from maintenance documents for technicians. At first, the idea seems straightforward: documents are stored in several places, some contain different version states, and, in the team's assessment, finding the right passage takes too much time. The first question is: “Which AI model can automatically summarize our maintenance documents?”

This question sounds like a procurement task. In fact, it already assumes several things: that a language model is the right means, that “summary” describes a sufficiently clear task, that the source documents are reliable and current, and that a shorter text makes searching better. None of these assumptions is confirmed by the question alone. Yet the wording already sets a direction: first find a model, then limit its use.

In an AI decision, this is an ethically relevant moment. Even before a system affects people, processes data, or intervenes in a process, the decision is made about what counts as a problem at all. Those who ask a question help determine which goals, alternatives, and harms will later be visible. If “save time” becomes the only goal, version reliability, traceability, working conditions, or the ability to spot an error in time may disappear from view.

The project-developed question expansion and prompt-optimization model F v5 starts with this prior decision. Its sequence A → B → D → C moves from input and context clarification through multiple hypotheses, meta-questions, and adaptive feedback to follow-up questions, fallback, iteration, and documentation. Only Module C consolidates the final adapted question as a working prompt, a transparent process summary, and a GROW action plan. [9] This sequence is ethically useful when it actually changes the initial question. A neatly formatted list of questions is not enough.

The thesis of this article is: F v5 can turn a premature tool request into a bounded, reviewable inquiry. The model produces neither truth nor fairness nor AI approval. A→B→D→C is the architecture of the project-developed model. The origins of individual components describe their purpose; they do not validate the combination or demonstrate better ethical AI decisions. That would require a predefined comparative evaluation in the relevant use context.

Questions are part of inquiry, not merely its packaging

John Dewey describes inquiry as work on an initially indeterminate situation: observations, concepts, hypotheses, and consequences are arranged so that a reasoned judgment becomes possible. This is a philosophical perspective on knowledge, not a modern AI governance procedure. It nevertheless helps avoid a common reduction. A question must not only be linguistically sound; it must fit the situation to be investigated. [1]

Ethics therefore does not begin only where a model produces possible harms. It also begins with the choice of what the team wants to measure, automate, delegate, or treat as “routine.” The questions “How do we generate summaries more quickly?” and “How can staff reliably find the currently approved manual passage?” do not describe the same task. The first privileges text production; the second makes access to information and the authority of the source the subject. A different question opens different solution options.

For social planning questions, Rittel and Webber showed that some problems cannot be treated like clearly bounded technical tasks: goals, affected interests, and the correct problem formulation itself may be contested. This does not mean that every operational problem is “wicked” or unsolvable. A wrongly named file path can probably be corrected directly. It becomes more difficult when several groups assess good performance, acceptable risk, or fair distribution differently. Then the search for a single neutral question can conceal the conflict instead of resolving it. [2]

The practical consequence is a distinction:

Routine question: Task, goal, and consequences are known; a brief review or direct action is proportionate.

Unclear decision question: Goal, terms, data, responsibilities, or success criteria are still open; the question must first be specified.

Value or governance conflict: Those affected weigh goals or error consequences differently; participation and a documented decision are also needed.

F fits primarily in the second category. It can also prepare material for the third. But it does not replace the consultation and governance process that includes different perspectives and establishes binding responsibilities.

The project’s own question model: from input to a reviewable working assignment

In what follows, F denotes the project-developed question expansion and prompt-optimization model F v5. Its sequence is A→B→D→C; in particular, D is addressed before C. This is not an editorial rearrangement into the more common A-B-C-D scheme. The loop keeps uncertainty, follow-up questions, and possible revision open before the final working assignment is formulated. The primary versions describe F as iterative work with user input and feedback, not merely as a facilitation diagram. [9]

The through-line is F v5 itself: input and context are clarified, hypotheses and feedback can change the working assignment, D keeps error and revision open, and C delivers a prompt, transparent rationale, and GROW plan. The techniques named in A support individual steps; they do not replace F v5. Socratic clarification, 5 Whys, SCAMPER, GROW, and Bloom each serve a defined purpose in the article. QFT is not applied; readers who want to study that external method separately can use the official description as optional further reading. [5][9]

A fully fictional SME case

The company in this example is entirely fictional. The scenario assumptions are: maintenance documents are kept in several storage locations; similar files have different version states; staff report search effort; there is no documented baseline. The team has not yet tested any model. There are no measured time savings or error rates, real pilot data, or experience reports.

Management proposes having a language model generate “summaries automatically.” Such wording mixes at least four possible tasks:

1. heavily shortening a document, 2. finding the passage in an authorized source, 3. formulating a new instruction from several sources, 4. giving a person an answer that they may follow immediately.

These tasks differ in risk. A search aid that points to an approved page is not equivalent to a generated repair instruction. Human review protects only when the reviewing person finds the right source, has the time and competence to check it, and is actually allowed to reject the output. For this case, that means: the first question must not simultaneously approve selection, testing, and use.

A: Clarify terms and cause before choosing a tool

The Socratic work here does not begin with a long philosophical debate. It asks several targeted follow-up questions: What does “summary” mean? Should it contain instructions or only show references? What does “current” mean? Who determines which version is valid? Which concrete activity is meant to become faster or more reliable? What would count as an error? What action should a person be able to take after the output? [8]

These questions already change the task. “Summarize automatically” is not a success criterion. A shorter amount of text may be faster to read while omitting a warning, condition, or version indication. A plausible answer may use the wrong file. An automatically generated formulation may also convey more authority than its source situation warrants.

In A, GROW is first used as a preliminary situation check and is brought together in C with the final question and next step. Goal: “Save time” is not yet a specific goal without measurement and an exact activity. Reality: In the scenario, there are several storage locations and unclear version states, but no baseline. Options: The initial request treats a model as the only option; an index, version register, or clearer document navigation are still missing. Will: A binding next action would be premature while it remains unclear which source is valid and who determines this. The first GROW pass therefore does not say “proceed,” but makes visible why there is not yet a sound action plan.

F-A begins with three questions developed for the case: Is the search difficulty caused by storage, terminology, or understanding? What must a person see to check the authority and version of a reference? What observation would justify comparing an AI search aid with a manual option at all? They follow the project model’s own work on input, goals, and assumptions; no external question sequence is applied. For a course exercise, Bloom can additionally name learning actions: participants first explain the difference between finding, summarizing, and instructing; then examine assumptions; and finally design a bounded question. Bloom classifies learning objectives; it does not establish question quality or truth. [9]

For clarification of causes, 5 Why can be used as a heuristic. Taiichi Ohno describes repeatedly asking why as part of problem exploration in the Toyota Production System. This does not mean that exactly five questions always uncover a true root cause. For social or organizational causes, a single chain can overlook important interactions, power issues, or alternative causes. Each answer is initially a testable hypothesis. [3]

The chain deliberately ends before an apparently final diagnosis. There may already be a responsible role, but no visible register. The document organization may be unproblematic and the search terms may not fit everyday work. The problem may occur only in a few unusual cases. F must not mistake a plausible explanation for an observed fact.

SCAMPER can then generate options, but not assess them. In the example, the question “What could we substitute, combine, or eliminate?” leads to several possibilities: automatic text production could be replaced by a versioned index; a search aid could display references and be combined with the original document; unreviewed step-by-step instructions could initially be excluded entirely from the scope of functions. Robert F. Eberle’s book documents SCAMPER as a procedure for generating ideas and variants. The variants still require evidence, safety, and governance review. [4]

Bloom has a different supporting role: it can classify learning actions. The 1956 taxonomy and its later revision are not identical; no taxonomy proves that a higher named level is automatically better or that a question is truer. [6]

An interim A result therefore looks like this:

Initial question: Which AI model can automatically summarize our maintenance documents?

A question: Which concrete search or information problem should be solved, which source counts as authorized, and why is AI assumed as the means for this?

This is a better working question, but not yet a decision. A has opened the search space and suspended premature model selection.

B: Let several explanations compete

A common misunderstanding is this: once the team has found the first plausible cause, it can move directly to the solution. B slows down this jump. It analyzes context, formulates several competing hypotheses, meta-questions, and alternative explanations, and integrates feedback to adjust the focus. The primary version proposes NLP-supported context analysis and illustrative confidence values; this article does not claim an implemented or evaluated NLP component or calibrated probabilities. The fictional case justifies its hypotheses qualitatively rather than presenting percentages or an intuition-based ranking. [9]

For the fictional company, at least four explanations remain possible:

A serious counter-hypothesis is that the reported delay does not occur often enough to justify a new system. The assumption “more AI brings more benefit” is not thereby replaced with “no AI is always better.” Instead, a comparison task arises: which concrete non-AI change solves the problem with less effort or fewer new risks?

Meta-questions examine how the team arrived at its question: Who formulated the assignment? Who defines a “correct” reference? Whose time saving counts? Who bears the consequences when the wrong page is used? Can staff point out an unclear or conflicting output without this being treated as resistance to technology? Which experience or language has so far been underrepresented in the discussion?

This makes a philosophical question practical: the question is not neutral when it writes certain interests into the goal and treats others as subsequent side conditions. The assignment “summarize more quickly” may be an organizational goal; it does not yet say whether the benefit is fairly distributed or an error would be tolerable. F can make these questions visible, but cannot itself determine how the organization should answer them.

Supporting techniques have limits too. SCAMPER can generate many unsuitable variants. A question-generation technique does not automatically make an omitted affected group visible. Bloom can classify a learning objective but cannot examine the facts. GROW structures an approach but cannot assess whether the goal is ethically defensible. The choice of every supporting tool therefore needs a reason: what should it do within F-A, and which question remains open afterward?

D: Take uncertainty, feedback, and fallback seriously

D is where a popular shortcut fails. An organization would like the questioning process to yield a clear final formulation. But if central terms, sources, or responsibilities remain contradictory, a smooth final question would be false clarity. D requires a follow-up question instead of filling the gap with a plausible assumption.

In the scenario, comparing the documents reveals that two documents differ in one maintenance step. This is a fictional development in the case, not a report about a real company. The new information changes what counts as the responsible next question. The team can no longer simply test which model produces the nicer short version. It must first clarify which source is valid and who bindingly determines that status. F v5 also assigns D the work of recommending external data sources, best practices, risk factors, and, where relevant, industry context. For this case, that means naming the authorized manual version, recording a manual baseline for the search task, and checking data/security risks as well as the risk of newly generated repair instructions. These are open checks, not established facts or a rollout recommendation. [9]

The D phase keeps at least four statuses distinct:

Unknown: It has not yet been measured how often the search fails or how long it takes.

Contradictory: In the scenario, there are different text versions for one step.

Normatively open: It has not been decided whether AI may ever generate action instructions.

Provisionally decided: Until clarification, no generated repair steps are issued to staff.

A fallback must work in practice. If documents conflict, the reference cannot be displayed, or a person cannot check the source, the example rule is: do not use the output; the responsible human role clarifies the source; until then, the original procedure remains in place. “Human in the loop” is not a safety argument as long as it is unclear whether a human can decide, intervene, and dissent in time.

Feedback explicitly returns to F-B. In a role exercise, a technical role invented for the case reports that “lead to the current manual passage” asks too little: without a document identifier, version state, and specific page number, the reference cannot be checked reliably. The feedback becomes new input to the hypothesis and question work; D documents what changes and checks fallback and open issues before C consolidates them. If new information arrives after an initial pass, B/D are iterated again; A→B→D→C is not reordered. Q4 now expressly requires a verifiable source reference; if it is missing, the fallback applies. This is a fictional instructional objection, not a quotation or report from a real course. Participation does not mean automatically adopting every piece of feedback, but it must be able to change the question, an option, or the next step.

The revision trail in the case is:

This revision does not claim that the fourth formulation is “objectively optimal.” It is better bounded: the result shows which parts of the first question were discarded, which assumptions were added, and which conditions were still kept open. Q4 may later also prove wrong or too broad. A good process documents what can reopen a question.

C: Final question and next step—still no solution

At the end of C, GROW is used as an action structure. The mapping is practical, not epistemic: GROW organizes a goal, the current situation, possible options, and the next will/action step. It demonstrates neither that the goal is legitimate nor that an option works. [7]